An Improved Multivariate Synchrosqueezing Wavelet Transform Denoising Method Using Subspace Projection

Peipei Cao,Kaijie Zhou,Huali Wang

doi:10.1109/access.2020.3007933

Abstract

In recent years, the development of multi-sensor has emphasized the need to directly process multi-channel (multivariate) data. In this paper, a novel multivariate synchrosqueezing wavelet transform denoising method combined with subspace projection (SWT-SP) is proposed. One of the key points of this method is to obtain an optimal orthogonal matrix which can project a multivariate observation signal to a signal subspace occupied by a clean signal and an orthogonal noise subspace occupied by noise. Furthermore, the high dimensional time-frequency representation based on the synchrosqueezing transform realizes the multichannel signal information fusion, and the subspace projection makes full use of the spatial diversity characteristics of the observed signal. Finally, signal energy produces the aggregation effect in the former dimension space, which improves the signal-to-noise ratio(SNR) of signals in the signal subspace. The performance of this algorithm for standard multichannel denoising is verified on both real-world data and synthetic signals. The reconstructed signal obtained the improvement of the highest SNR by about 6 dB under different conditions.

Highlights

Recent advances in multi-sensor and data acquisition technology have facilitated routine recordings of multivariate or multichannel data sets, e.g., electroencephalogram (EEG), sofar signal [1]
We mainly introduce a multivariate synchrosqueezing wavelet transform denoising method combined with a subspace projection (SWT-SP)
EXPERIMENTAL SIMULATION RESULTS AND DISCUSSION To test the denoising performance of the proposed multivariate synchrosqueezing wavelet denoising combined with the subspace projection method (SWT-SP), simulation experiments are carried out on both synthetic data and real data respectively

Summary

INTRODUCTION

Recent advances in multi-sensor and data acquisition technology have facilitated routine recordings of multivariate or multichannel data sets, e.g., electroencephalogram (EEG), sofar signal [1]. A multivariate extension of EMD, namely multivariate empirical mode decomposition (MEMD) was proposed in [11] This method is limited to the univariate signal under Gaussian noise. The computation of synchrosqueezing transform based on wavelet is very time-consuming which not fits for the real-time signal denoising To resolve this issue, Ahrabian and Mandic proposed a time-frequency representation of high-dimensional signals by using synchrosqueezing wavelet transform (SWT) [14], and further realized the joint noise suppression of multi-channel signals. The SWT method can obtain the best noise suppression performance based on wavelet transform, it is only adaptive to the signal with slow frequency change, so its applications are limited and it is time-consuming. We mainly introduce a multivariate synchrosqueezing wavelet transform denoising method combined with a subspace projection (SWT-SP).

BASIC PRINCIPLES OF SWT

EXPERIMENTAL SIMULATION RESULTS AND DISCUSSION

CONCLUSION